A Linear DBSCAN Algorithm Based on LSH

• Published in: 2007 International Conference on Machine Learning and Cybernetics Vol. 5; pp. 2608 - 2614
• Main Authors: Yi-Pu Wu, Jin-Jiang Guo, Xue-Jie Zhang
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.08.2007
• Subjects: Clustering; Clustering algorithms; Computer science; Cybernetics; DBSCAN; Image segmentation; Information retrieval; Large-scale data; Large-scale systems; LSH; Machine learning; Machine learning algorithms; Nearest neighbor searches; Partitioning algorithms; Unsupervised learning
• ISBN: 1424409721; 9781424409723
• ISSN: 2160-133X
• DOI: 10.1109/ICMLC.2007.4370588

DBSCAN algorithm is used widely because it can effectively handle noise points and deal with data of any type in clustering. However, it has two inherent limitations: high time complexity O(NlogN) and poor ability in dealing large-scale data. In this paper, a linear DBSCAN based on LSH is proposed. In our algorithm the process of Nearest Neighbor Search is optimized by hashing. Compared with the original DBSCAN algorithm, the time complexity of this improved DBSCAN descends to O(N). Experimentally, this improved DBSCAN makes a significant decrease in the running time while maintaining the Cluster quality of the results. Moreover, the speedup (the running time of original DBSCAN algorithm divided by the running time of improved algorithm) increases with the size and dimension of dataset, and the parameter Eps of our algorithm does not have a strong influence on the clustering result. These improved properties enable DBSCAN to be used in a large scope.